| Seminar: |
CPS Seminar
|
| date: |
2022 October 28 (FRI) 16:00-18:00
|
| Room: |
Online seminar (Zoom)
|
| Speaker: |
ENDO Katsuhiro (Resercher,National Institute of Advanced Industrial Science and Technology
AIST)
|
| Title: |
"PIKG/MN-Core: Compier for particle-particle interaction calculation on
MN-Core"
|
| Abstract: |
MN-Core, a processor for deep learning jointly developed by Kobe University and PFN, supports double-precision operations even though it is designed for deep learning, and has high potential for general-purpose computation. It won first place in the Green 500 three times. However, until now, its software development environment has been available only for deep learning, and it has not been clear how it can be used in other HPC applications. We have extended PIKG, the DSL we have developed for describing particle-particle interactions, to generate code optimized for MN-Core. MN-Core processor can take the local memory as its operands, and placing intermediate data in appropriate storage elements is key to achieving high execution efficiency. We optimized this placement using simulated annealing to achieve performance equivalent to human placement. We believe that the success of this code generation and optimization has made it possible to realize general numerical computation on MN-Core. |
| Keywords: |
HPC, particle-particle interaction, Compiler, MN-Core
|
| Organizer: |
MAKINO Junichiro
|
